Carried forward from the #1907 perf investigation (four passes, three board A/B rounds; full accounting in plans/gfql-benchmark-numbers/plan.md). Neither is release-blocking; both are honest debits.
1. pandas @100k carries ~+2% vs the pre-#1905 floor. Pooled per-iteration p50: floor 89.76 / prev-head 90.99 / current 91.75 on q8. This is the trail correction's absolute cost (~2.2 ms probe, _two_hop_trail_illegal_pairs) landing on a fast cell. Blocking it for pandas was measured and is worse (+1.84 ms, significant — per-block call overhead exceeds the pass saved), and five probe formulations cluster within 0.26 ms, so the pass itself is at its floor. The removable route is a soundness proof rather than a faster scan: consult a degree fact for self_loops == 0 on the matching partition ⇒ zero illegal pairs, skipping the scan entirely. Index-dependent, so it cannot fire on the board's --binding cold arm, but it would help indexed users.
2. polars q8 @20k nominally +6.7% over floor (2.39 vs 2.24 ms), p=0.11 — not resolvable at 18 slots/build. Either confirm with more slots or drop the claim; do not quote a 20k q8 delta in either direction as things stand.
3. Harness placement. The per-query harnesses written for this investigation (in-situ variant swap, lane instrumentation, formulation shoot-outs) live in a session scratchpad. Per the standing convention that benchmark CODE lives in pyg-bench and pygraphistry carries only results, they should move to pyg-bench if these measurements are meant to be durable.
Context for anyone picking this up: the residual cost IS the correctness fix — the count is one bandwidth-bound re-read of both endpoint columns (five formulations within 0.126 ms, ~19.4 MB at ~16 GB/s). The only two ways to skip it are DegreeFact.self_loops (O(1), needs an index) and the bounds-proof by-product now used on the polars dense lane.
Carried forward from the #1907 perf investigation (four passes, three board A/B rounds; full accounting in
plans/gfql-benchmark-numbers/plan.md). Neither is release-blocking; both are honest debits.1. pandas @100k carries ~+2% vs the pre-#1905 floor. Pooled per-iteration p50: floor 89.76 / prev-head 90.99 / current 91.75 on q8. This is the trail correction's absolute cost (~2.2 ms probe,
_two_hop_trail_illegal_pairs) landing on a fast cell. Blocking it for pandas was measured and is worse (+1.84 ms, significant — per-block call overhead exceeds the pass saved), and five probe formulations cluster within 0.26 ms, so the pass itself is at its floor. The removable route is a soundness proof rather than a faster scan: consult a degree fact forself_loops == 0on the matching partition ⇒ zero illegal pairs, skipping the scan entirely. Index-dependent, so it cannot fire on the board's--binding coldarm, but it would help indexed users.2. polars q8 @20k nominally +6.7% over floor (2.39 vs 2.24 ms), p=0.11 — not resolvable at 18 slots/build. Either confirm with more slots or drop the claim; do not quote a 20k q8 delta in either direction as things stand.
3. Harness placement. The per-query harnesses written for this investigation (in-situ variant swap, lane instrumentation, formulation shoot-outs) live in a session scratchpad. Per the standing convention that benchmark CODE lives in pyg-bench and pygraphistry carries only results, they should move to pyg-bench if these measurements are meant to be durable.
Context for anyone picking this up: the residual cost IS the correctness fix — the count is one bandwidth-bound re-read of both endpoint columns (five formulations within 0.126 ms, ~19.4 MB at ~16 GB/s). The only two ways to skip it are
DegreeFact.self_loops(O(1), needs an index) and the bounds-proof by-product now used on the polars dense lane.